Anthropic makes Claude Code Auto Mode default August 14 with safety classifier

Anthropic will flip Auto Mode on by default for Claude Code across Pro, Max, and Team plans starting August 14, removing per-action permission prompts in favor of a background AI classifier that evaluates each proposed action and blocks destructive or out-of-bounds ones. Anthropic cites internal data showing the classifier catches 89% of dangerous commands compared to just 13.6% for human reviewers — a striking claim that reframes automated oversight as safer than manual approval.

Mechanically, the classifier runs as a background evaluator on the agent's tool calls, intercepting commands like destructive file operations or out-of-scope network actions before execution. Anthropic says it will not charge users for the token overhead the classifier consumes, removing a friction point for adoption. Claude Code lead Boris Cherny said he 'couldn't imagine going back to permission prompts.'

The developer community split sharply. Many cited the 89% vs 13.6% catch rate approvingly ('data doesn't lie'), while others fixated on the 11% slip-through rate and the philosophical shift of ceding approval authority to another model. The move also lands against a backdrop of prompt-injection anxiety — Cherny himself posted a widely-shared warning that malicious website text remains the most common attack vector against agents.

Competitively, this pushes Anthropic ahead on agent autonomy defaults just as AWS ships Dogwood temporal-policy governance and OpenAI grapples with agent breakout risks. Skeptics note the 89% figure is Anthropic's own internal measurement with no independent audit. What to watch: real-world incident reports after August 14, whether the classifier's false-positive rate frustrates power users, and whether competitors adopt similar always-on classifiers as the norm.